Dynamics of microseismicity and its relationship with the active structures in the western Corinth Rift (Greece)

HAL Id: hal-01845974

https://hal.archives-ouvertes.fr/hal-01845974

Submitted on 22 Jan 2019

HAL is a multi-disciplinary open access

archive for the deposit and dissemination of

sci-entific research documents, whether they are

pub-lished or not. The documents may come from

teaching and research institutions in France or

abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est

destinée au dépôt et à la diffusion de documents

scientifiques de niveau recherche, publiés ou non,

émanant des établissements d’enseignement et de

recherche français ou étrangers, des laboratoires

publics ou privés.

To cite this version:

Clara Duverger, Sophie Lambotte, Pascal Bernard, H. Lyon-Caen, Anne Deschamps, et al..

Dy-namics of microseismicity and its relationship with the active structures in the western Corinth Rift

(Greece). Geophysical Journal International, Oxford University Press (OUP), 2018, 215 (1), pp.196

-221. �10.1093/gji/ggy264�. �hal-01845974�

GJI Seismology

Dynamics of microseismicity and its relationship with the active

structures in the western Corinth Rift (Greece)

C. Duverger,

S. Lambotte,

P. Bernard,

H. Lyon-Caen,

A. Deschamps

and

A. Nercessian

1_{Institut de Physique du Globe de Paris, Sorbonne Paris Cit´e, Universit´e Paris Diderot, CNRS, 75005 Paris, France. E-mail:clara.duverger.pro@gmail.com} 2_{Institut de Physique du Globe de Strasbourg, UMR7516, Universit´e de Strasbourg, EOST/CNRS, 67000 Strasbourg, France}

3_{Laboratoire de G´eologie, ´Ecole Normale Sup´erieure de Paris/CNRS UMR8538, PSL Research University, 75005 Paris, France} 4_{Universit´e Cˆote d’Azur, CNRS, Observatoire de la Cˆote d’Azur, IRD, G´eoazur, 06560 Valbonne Sophia Antipolis, France}

Accepted 2018 June 28. Received 2018 February 9; in original form 2018 June 26

S U M M A R Y

We analyse the complete earthquake archive of the western Corinth Rift using both cross-correlations between pairs of event waveforms and accurate differential traveltimes observed at common stations, in order to identify small-scale fault structures at depth. The waveform database was generated by the dense Corinth Rift Laboratory network and includes about 205 000 events between 2000 and 2015. Half of them are accurately relocated using double-difference techniques. The novelty of this relocated catalogue is the integration of the recent westernmost earthquakes due to the extension of the network in 2010 to the western extremity of the Corinth Rift and the consideration of the whole database over more than 15 yr. The total relocated seismicity exhibits well-defined clusters at the root of the main normal faults mainly between 5 and 10 km depth in the middle of the gulf and illuminates thin active structure planes dipping north about 20◦under the northern coast. Some seismicity is observed in the footwall of the main active faults, along the West and East Helike faults. We also built a multiplet database based on waveform similarity taking into account cross-correlation coefficients weighted by signal-to-noise ratios. Short-term multiplets are concentrated in the middle of the gulf along the Kamarai fault system, in a 1–2 km thick layer at 6–8 km depth, interpreted as a highly fractured geological layer. They are often associated to slow seismic migration velocities occurring in this zone during strong swarm episodes and are thus likely to be triggered by pore pressure variations. On the other hand, most long-term and regular multiplets are located deeper (7– 10 km), under the northern coast, within a layer less than 0.3 km thick. They occur at the border of nearly planar structures with low seismicity rate, which we identify as fault planes, and they may be explained by aseismic slip on the fault surface around them. This supports the existence of an immature structure growing downdip towards the north at the base of the active geological layer, which possibly connects to the ductile middle crust around 15 km depth, as suggested by the occurrence of deeper events in the continuity of the 1995–fault plane. The different migration velocities (from 0.05 km d−1 to several km d−1) highlighted during the western 2014–swarms indicate that both pore pressure and creep diffusion are operating in the fault zone. The fast migrations observed in the Psathopyrgos fault zone, where a slow slip event was detected by dilatometers in 2002, compare with that for creeping faults. To the west, from spatial distribution of events, we show that the Rion–Patras fault connecting the western extremity of the Corinth Rift fault system to the Patras Rift, is dipping around 60◦north–west with a rake angle of−115◦. Finally, we identified two new areas within the central active zone which may correspond to large scale, locked asperities on active fault surfaces, similar in size to the main asperity broken during the 1995, MW6.3, Aigion earthquake.

Key words: Europe; Transient deformation; Earthquake dynamics; Seismicity and tectonics.

196 C_{The Author(s) 2018. Published by Oxford University Press on behalf of The Royal Astronomical Society.}

extending E–W over 100 km, where south coast uplifts of several millimetres per year over at least the last 0.3 Myr (e.g. Vita-Finzi

1993; Armijo et al.1996; Rohais2007). The fault system (Fig.1) is mainly composed of onshore antithetic normal faults, with the main active faults cropping out on the southern coast and dipping north at 40◦–60◦. Some active offshore faults have also been identified in the middle of the gulf (e.g. Stefanos et al.2002; Moretti et al.

2003; Bell et al.2008; Beckers et al.2015). On the northern coast, the south dipping normal faults have been less extensively studied. Currently, the north–south extension of the rift is inhomogeneous, with a high extension rate of around 15 mm yr−1in the western part and 10 mm yr−1in the eastern part (Briole et al.2000; Avallone

et al.2004). The Corinth Rift has one of the highest strain rates in the world (about 10−15s−1). This deformation is the consequence of the backarc extension due to the Hellenic subduction in the south and the westward propagation of the North Anatolian Fault (Armijo

et al.1996). In the history of the rift activity, the older inland normal faults have been progressively deactivated and its deformation has been translated northwards, presently centreed in the middle of the gulf. The western tip of the gulf has shown evidence of strike-slip motion (Rigo et al.1996; Beckers et al.2015).

From a geological point of view, the first 5–10 km of the crust is made of the Hellenide nappe stack, composed of a succession of five different carbonate nappes. Three of them outcrop in the western and central Corinth Rift (Fig.1). These geological struc-tures are inherited from the past east–west convergence during the Miocene. They may have influenced the sedimentation and the fault segmentation (Ghisetti & Vezzani2004) and may play a role in the lateral variability of the seismicity distribution (Gautier et al.2006; Lambotte et al.2014; Duverger et al.2015).

The Corinth Rift has been shaken by a few destructive M> 6 earthquakes per century, making it one of the most seismically active regions in Europe. For instance, in the 19th century, two historical earthquakes that occurred in 1817 and 1888 have been associated to the Aigion fault and one in 1861 to the Helike fault (Albini et al.2017; Fig.1) . However, some large recent earthquakes of the region do not seem to be associated to one of the main fault planes visible at the surface. This is especially the case for the major earthquake MW6.3 in 1995, which occurred on a low-angle north

dipping blind fault and caused severe damages in the city of Aigion (Bernard et al.1997). The analysis of the doublet of magnitude

MW> 5 earthquakes in 2010 emphasizes a link with shallow steep

dipping fault structures, but has not been associated to any known major normal faults (Sokos et al.2012). The fault plane extensions at depth are not well constrained or sometimes totally unknown, so that linking an earthquake to a fault remains complicated even nowadays.

The microseismic activity follows a swarm organization, with alternation of intensive crisis and more quiescent periods. The con-trolling factor of these seismic swarms is still not well understood, but some crises have been explained by fluid circulations or pore pressure diffusions (Bourouis & Cornet2009; Pacchiani & Lyon-Caen2010; Duverger et al.2015). The microseismicity is mainly concentrated into a 3 km thick layer below the gulf, slightly dipping

to root into this layer, since almost no seismicity is observed below 9 km depth beneath this layer.

Using more than 20 000 events of moment magnitudes ranging from 1 to 4, the objectives of this paper are (1) to present new insights into the westernmost deep structures and their relation-ships with the faults observed at the surface, (2) to characterize the spatio-temporal evolution of the microseismicity and (3) to identify the interactions between seismic and aseismic mechanisms. With this aim in mind, we perform waveform cross-correlations and a double-difference relocation coupled with a multiplet classification to identify small-scale structures within the microseismicity and to characterize their reactivation through years. Thanks to the west-ward extension of the seismological network, we extend the study of Lambotte et al. (2014) to the westernmost part of the Corinth Rift and perform an homogeneous and detailed analysis of the 2000– 2015 data. Recently, Mesimeri et al. (2017) presented a relocation of about 30 000 events in the entire Corinth Rift from 2008 to 2014 using the regional HUSN catalogue. In our paper, we present a relo-cation of the whole 16 yr available data using the dense CRL local network and a more complete catalogue (completeness magnitude

MW∼ 1.2). In addition, this large catalogue allows us to precisely

characterize the temporal evolution of the seismicity and investigate its dynamic implications for the deformation in relationship with the detailed structures of the fault system and for the seismic hazard. After a description of the local seismic network and data sets used, we explain the different steps of the relocation process and the multiplet extraction. Then we present the high-resolution images of the global relocated microseismicity, describing the fault geometries at depth. We statistically analyse the spatio-temporal evolution of the multiplets to propose different mechanisms or stress responses across the rift. We finally discuss the active fault structures in terms of possible mechanical sources and seismic hazard.

2 D AT A A N D M E T H O D S

We take advantage of the dense distribution of recorded events within the western Corinth Rift over the last two decades to obtain a high-resolution image of the microseismicity at depth. We proceed in three main steps: (1) cross-correlation of waveforms, (2) relo-cation of the seismicity using a double-difference technique and (3) identification of multiplets by hierarchical clustering. Here, we describe the data set used and the different stages of the processes.

2.1 Seismic network and waveform data set

We use the entire seismic archive, 16 yr of data recorded by the CRL network (Corinth Rift Laboratory,2013), from spring 2000 to winter 2015. This database includes around 225 000 events, slightly less than 14 million digital waveforms (800 Gb) recorded at about 15 different permanent or temporary stations on an average. The first permanent stations of the CRL network have been installed in 2000 in the western part of the rift around the city of Aigion by a French CNRS team. Between 2000 and 2009, the network

Figure 1. Tectonic map of the central and western part of the Corinth Rift. Onshore active faults are from Ford et al. (2009,2013); Palyvos et al. (2005,2008), main offshore faults are from Flott´e et al. (2005); Bell et al. (2009); Taylor et al. (2011); Beckers et al. (2015). Isobaths are from Bell et al. (2008) and GPS measurements are from Avallone et al. (2004). Epicentres and focal mechanisms of the two MW5 earthquakes are from Sokos et al. (2012) and from Bernard et al. (1997) for the MW6.3 Aigion earthquake. Historical earthquakes of magnitude M≥ 6 affecting faults in the study area (double thick lines) are mentioned by dates (Albini et al.2017). The distribution of the onshore Hellenide nappes and the Plio–Pleistocene sediments are drawn and described in the legend. The insert shows the Hellenic subduction in the south and the dextral strike-slip North Anatolian Fault. Modified from Beckers et al. (2015).

covers a 30 km× 30 km area with 11 permanent stations recording at 125 Hz, equipped with three-component seismometers (ALIK, AGEO, AIOA, DIMT, ELE, KOU, PANR, PSAR, PYRG, TEME and TRIZ). From 2010, four new three-component velocimeters recording at 100 Hz have been installed further west to especially monitor the activity of Psathopyrgos zone (AGRP, MALA, ROD3 and ZIRI). Some temporary stations (FIL, MYL and HELI) have been installed to monitor aftershock activity in 2010 on the northern coast and a large swarm activity in 2013 on the southern coast. In 2013, an antenna with seven broad-band stations has been deployed in Magoula (MG0). Nowadays, the CRL network covers a 60 km× 40 km area (Fig.2).

The CRLNET also includes seven stations recording at 100 Hz equipped with broad-band seismometers operated by the University of Athens (KALE and LAKA) and the University of Patras (ANX, EFP, SERG, UPR and VVK). Since 2010 all stations are recording at 100 Hz, we thus down-sampled all previous records at 100 Hz to homogenize the database.

A post-trigger algorithm, based on an STA/LTA (short-term average/long-term average) ratio, is used to extract events from the continuous CRL records. It detects an event when more than three stations trigger at about the same time (within 3 s) and picks the P and S phases when possible. Details on the detection and picking algorithms are available in Bourouis & Cornet (2009). All events are automatically located using hypo71 software (Lee & Lahr 1972), the 1-D velocity model of Rigo et al. (1996) and a

VP/VS ratio of 1.80. The moment magnitude of events has been

computed by spectral inversion of the seismograms (Satriano et al.

2016,https://gitlab.com/claudiodsf/sourcespec). Part of the data set (around 10 per cent) has been manually picked for specific studies. In this case, we prefer to use manual picks than automatic ones. Inaccuracies in the phase picks and errors in the initial model affect the hypocentre location errors. These latter vary from several hun-dreds metres to a few kilometres for events in the area covered by the network, with depths less well constrained than epicentres. The average standard errors for events outside the network can reach 5 km (Lambotte et al.2014). These errors are many times larger than the spatial dimension of microearthquakes themselves. Indeed, magnitude 1–4 earthquakes have source dimension typically rang-ing from 10 to 1000 m. Thus, studyrang-ing the seismic hazard of active faults or the physical properties of these earthquakes is limited by these location uncertainties.

2.2 Computation of waveform cross-correlations

Here, we take advantage of the density of events recorded in the Corinth Rift. A common signal processing tool, the cross-correlation is used both (1) to identify multiplets (a multiplet is a group of similar events, spatially close with identical focal mech-anisms) and (2) to perform a relocation at local scale using the dif-ferential phase traveltimes between earthquakes (e.g. Waldhauser & Schaff2008). Such differential times are simultaneously inverted for estimating more precisely the distance between events (e.g. Got

et al.1994; Waldhauser & Ellsworth2000).

Figure 2. Map of the seismic stations used for the cross-correlations and the relocation. Black and white triangles (CL) are CNRS stations. Among them, the easternmost stations (ALIK, AGEO, AIOA, DIMT, ELE, KOU, PANR, PSAR, PYRG, TEME and TRIZ) started recording in 2000–2001 but ELE was removed in 2006 and KOU in 2011. FIL and MYL were temporary stations used in 2010 to record aftershocks of the two MW5 earthquakes occurring in January 2010 on the northern coast (see Fig.1). HELI is also a temporary station installed later in 2013 to record a large swarm occurring on the southern coast. ROD3 and ZIRI started recording in 2010, AGRP and MALA in 2012. MG0 (black square) is a CNRS antenna with seven broad-band seismometers set up in 2013. Light grey triangles (HP) are stations operated by the University of Patras (one of these stations, ANX, is off-map, 15 km just north of EFP). Dark grey triangles (HA) are stations from the University of Athens.

We use the time-domain cross-correlation to estimate the corre-lation coefficient CCwhich quantifies the similarity between two

event waveforms recorded at a same station, and the time-shift

τ which only corresponds to the traveltime difference if the initial

alignment of the waveforms (zero lag) is at the corresponding phase arrival-time picks. Measuring traveltime differences and similarity coefficients by cross-correlation requires the choice of windows around phase arrival times for a pair of events. A battery of tests were conducted by Lambotte et al. (2014) to define correlation pa-rameters that produce robust delay time measurements in an efficient manner for the Corinth Rift. Based on these tests and some read-justments linked to the evolution of the CRL network from 2010, we choose the following input parameters to run across the entire network.

Seismograms were filtered between 1.5 and 15 Hz, instead of 3 and 20 Hz as in Lambotte et al. (2014), in order to allow for cor-relations between larger events (MW≥ 3). According to Schaff &

Waldhauser (2005) who studied California data with a similar size problem as ours, filtering from 1.5 and 15 Hz increases the num-ber of useful cross-correlation measurements reducing long-period instrument noise and less similar high frequencies. Correlation mea-surements were made on 1 s time-window lengths for the P-wave train and 2 s for the S-wave train, giving the opportunity to look for time-shifts up to 0.5 s for the P- and 1 s for S-windows. The waveform windows begin at 0.3 s and 0.5 s before the P- and S-pick

arrival time, respectively. The P-windows were extracted from the vertical component and the S-windows were computed on the two horizontal components (east and north).

Cross-correlations are first computed at the time resolution of one sample (0.01 s) by aligning to the nearest sample, and fi-nally at sub-sample time resolution using a polynomial fit in the vicinity of the peak of the cross-correlation function. This proce-dure enables the measurement of delays within millisecond pre-cision and allows a reliable recovering for large delays (Schaff

et al.2004).

Numerically cross-correlating all events with each other in the case of more than 200 000 events in the database, would be hardly feasible because of the size of the problem. To reduce it and for the purpose of the relocation procedure described in the Supplementary Information SA.1.2, we first divide the western Corinth Rift into 10 rectangles (Supplementary Information Fig. SA.2). The edges of boxes were chosen to lie as much as possible in regions of sparse seismicity and they overlap to ensure the spatial continuity in the seismicity.

In order to reduce the computational time, we do not cross-correlate every event with every other event of the same area, but only event pairs less than 5 km apart. This radius was chosen taking into account the preliminary hypocentre location uncertainty and the known degradation of the waveform similarity with interevent separation distance.

A total of 15 billion cross-correlation measurements (P waves on vertical component and S waves on the two horizontal components) were performed. The computations were carried out on 10 nodes of the S-CAPAD1 _{cluster at a rate of about 30 million measurements}

per CPU hour.

2.3 Double-difference relocation

The resulting cross-correlation time-shifts are used to perform a high-precision relocation using double-differences of traveltimes (e.g. Got et al.1994; Hauksson & Shearer2005; Waldhauser & Schaff 2008), with HypoDD software (Waldhauser & Ellsworth

2000). All the details about the data selection and weighting, the pa-rameters for the inversion and the computational strategy are given in the Supporting Information SA.1. We also provide statistics on cross-correlation data in the Supporting Information SB, particu-larly the influence of signal-to-noise ratio (SNR; Supplementary Information Fig. SB.3), distance between events (Supplementary Information Fig. SB.4) and magnitude difference (Supplementary Information Fig. SB.5) on correlation measurements. Statistics on cross-correlations for P- or S-wave windows and noise windows are also presented in Supplementary Information Figs SB.1 and SB.2, respectively.

In total, we relocate around 94 000 events between 2000 and 2015 in the western Corinth Rift, from the 115 000 selected ones which were recorded by four or more stations with at least six picked phases (Fig.3). The new hypocentres show improved clustering both horizontally and vertically, creating a more focused picture of the complex distribution of seismicity. The HypoDD program does not estimate hypocentral errors directly and it was not feasible to carry out detailed error estimates here because of the large data set size. Nevertheless, some statistical tests using a bootstrap technique have previously been done on a smaller data set by Lambotte et al. (2014) to assess the relative relocation errors. According to the authors, which used the same relocation approach, the maximum relative relocation uncertainty is about 30 m in the centre of the rift where the density of events and the quantity of cross-correlations is high, and can reach few hundreds of metres at the edges where the network coverage is not sufficient.

2.4 Classification of earthquake multiplets

The resulting cross-correlation coefficients are then used to extract multiplets. Multiplets can provide additional information on the geometry of microstructures and are useful to characterize seis-mically active structures (e.g. Rietbrock et al.1996; Pacchiani & Lyon-Caen2010). The whole methodology used to estimate a rep-resentative similarity coefficient (Sij) between two events, i and j,

and define families of events is described in the Supplementary Information SA.2.

This Sij(Supplementary Information eq. SE.2) value ranges

be-tween 0 and 1 and is computed taking into account the cross-correlation coefficients of the P and the S phase at each common station for a pair of events, weighted by the reduced SNRs of the waveforms (Supplementary Information eq. SE.1). The main par-ticularity of our method is the use of SNRs of the body phases to weight the data. Basically, this weight increases nonlinearly from 0 to 1 with increasing SNR (Supplementary Information Fig. SA.1)

1_{Service de CAlcul PArall`ele et de traitement de Donn´ees en sciences de la} Terre

and depends on the pick type (automatic or manual). Data with SNR<1.5 are given a weight close to 0, whereas data with SNR

>5 are given a weight close to 1. It appears to be a robust metric

distance to classify events in multiplets. Their extraction is con-ducted by hierarchical clustering technique, which is detailed in the Supplementary Information SA.2.

Finally, we choose a similarity cut-off Sc = 0.8 to define the

multiplets and perform the structural analysis. The following de-scription of families of events will then correspond to this threshold value. In this case, 70 per cent of the events of the CRL database are classified in multiplets, among those 66 per cent are in doublets.

In order to distinguish repeaters (a repeater is a group of events rupturing systematically a same asperity) from multiplets, we con-duct a more detailed analysis with a stronger threshold. Among multiplets obtained with a 0.8 similarity cut-off, we cross-correlate their respective events with a 5 s waveform window encompassing the P and the Sphase together. Indeed, repeater events should pre-serve the differential time between the P and S phase apart from the high waveform similarities. We then apply a higher similarity cut-off Sc= 0.9 to construct repeaters.

3 R E S U L T S

After the relocation stage and the multiplet extraction, we obtain a high-precision image of the microseismicity at depth with the clustering of families of similar earthquakes. The Fig.3b presents the entire western Corinth Rift relocation map over 16 yr. We clearly observe that the seismicity is not distributed homogeneously across the gulf.

For the presentation of the results, we divided the studied area in three main sectors: the central zone (ZC), the western zone (ZW) and the westernmost zone (ZM), which correspond to different seismic rate and spatial distribution of the microseismicity (Fig.4a). We also introduced some acronyms for specific structures which will be described in this section: the North Eratini zone (NEZ), the Transition zone (TZ), the Trizonia Island zone (IZ), the Aigion-Fassouleika fault zone (AFZ) and the Rion-Patras and Psathopyrgos fault zone (RPZ).

To describe these structures, we present eight vertical cross-sections along the rift in the Fig.5and a dozen more in the Sup-plementary Information Fig. SC.1. The spatial extension of these cross-sections is shown in the Fig.4b.

3.1 Large-scale microseismicity patterns

The absence of shallow events makes difficult to evaluate the geom-etry of faults at depth and to constrain their dips. However, analysing thinly clustered events in several cross-sections helps to define and interpret their geometry.

3.1.1 The central part (ZC) versus the western part (ZW)

The western part (ZW) concentrates in average more than 10 events per day since 2000 and up to hundreds of events during swarms. Its seismicity is mainly located in the middle of the rift, offshore (Fig.4), between 4 and 10 km depth, in a 1–2 km thick layer gently dipping north (sections P7 and P8, Figs5c-d). Seismicity structures are complex and highly variable over short distances.

On the contrary, the ZC exhibits a lower seismicity rate with a seismic gap in the middle of the rift (Fig.4and section P4, Fig.5a). Few small swarms are located under the northern coast at a depth

Figure 3. Maps of initial and relocated earthquakes. (a) Map of selected earthquakes for the relocation (n= 115 000) from 2000 to 2015, recorded by four or more stations (triangles) with at least six picked phases. These events are plotted at their original location obtained with hypo71. (b) Map of the relocated earthquakes (n= 94 000) computed with hypoDD. Dot size is proportional to the magnitude. Colours represent the depth of events. Black lines are the main normal faults of the region reported in Beckers (2015).

larger than 10 km, in a less than 1 km thick layer, dipping 25◦N (NEZ). The seismicity continues deeper further north, reaching al-most 15 km depth. The fault system of ZC is shifted to the north compared to ZW, and its seismicity is deeper. This clear disconti-nuity in the seismic activity and the tectonic settings occurs over a

small distance range (5 km) and will be described in more details as TZ in the Section 3.2.

The northwesternmost seismicity of the ZW mainly corresponds to the aftershock sequence after a doublet (MW5.3 and 5.2)

oc-curring in 2010 January 18 and 2010 January 22 (zone 2AZ on

Figure 4. Main structures discussed and clusters delineated by the relocated seismicity. (a) Location of the different sectors (red dashed lines) and the main structures / clusters (blue ellipses) discussed in this section. Black lines are the main normal faults of the region reported in Beckers (2015). Grey dots are the relocated earthquakes. Black triangles indicate the seismic stations. White stars are the same MW≥ 5 earthquakes as those of Fig.1. Initials are associated to names defined and used in the text (ZC, Central rift Zone; ZW, Western Zone; ZM, Westernmost Zone; NEZ, North Eratini Zone; TZ, Transition Zone; IZ, Trizonia Island Zone; AFZ, Aigion–Fasouleika Fault Zone; 2AZ, Aftershock Zone of the 2010 doublet; RPZ, Rion–Patras—Psathopyrgos Fault Zone). (b) Location of the cross-sections detailed in Fig.5for the thick red rectangular boxes, and in the Supplementary Information Fig. SC.1 for the regular blue rectangular boxes.

the Fig.4a), close to the Efpalio city (Sokos et al.2012; Ganas

et al.2013). These events are the largest occurring in the western Corinth Rift since the 1995, MW6.3 Aigion earthquake.

Hypocen-tres of the first and the second major events are respectively re-located at (38.428◦ N, 21.890◦) at 11.5 km depth and (38.441◦, 21.964◦) at 10.5 km depth (Fig.4a and Supplementary Information Fig. SC.1.lm). Our epicentre locations are in agreement with the

previous estimates reported and compiled in the table 1 of Ganas

et al. (2013). Our depth estimate of the January 18th event is among the deepest proposed in other studies, which are ranging from 2 to 12 km. The constant pattern retrieved over all studies is that the second main event is systematically shallower than the first one. According to Sokos et al. (2012), the rupture of the first event most likely occurred on a 55◦blind south-dipping nodal plane whereas its

Figure 5. Vertical cross-sections along the western Corinth Rift. They are identified by thick red rectangles in Fig.4b and numbers. Bold black lines indicate the major onshore and offshore faults, with dashed thick lines for their extrapolation at depth. Light blue lines at the surface represent the offshore rift. Black triangles at the surface are the seismic stations. Dark blue lines at panel bottom represent the extension of zones defined in Fig.4a and described in the text. The colour scale and the dot size are proportional to the moment magnitude of events. The white star in (a) is the location of the MW6.3 1995 earthquake, which ruptured a blind fault plane dipping∼30◦and drawn in bold black line. The white star in (e) locates the MW5.3 first major event of the doublet occurred in January 2010 (see Fig.1). The zoom in (c) highlights some antithetic structures sub-parallel to the Trizonia fault plane, thanks to event clusters occurred in summer 2014.

Figure 5. continued.

sequence. Our results support these observations as our relocated aftershocks stay confined in the same spatial area.

3.1.2 The westernmost part (ZM) versus the western part (ZW)

In the westernmost part of the rift (ZM), the seismicity seems more clustered than in the ZW, where larger and laterally extended swarms are observed. This visual impression has to be tempered by the fact that no station was installed before 2010 in the ZM. If we only consider data above the completeness moment magnitude MWc =

1.2, from 2011 when the network covers the full area, we do not observe denser and larger spatially continuous swarms in the ZW than in the ZM, but sequences of small clusters detached from each other.

Clusters in the ZM are associated to the roots of the E–W Psathopyrgos fault and the NE–SW Rion–Patras fault, which follow the south coastline (Fig.3). The lateral extension of clusters is also parallel to the two faults. The seismicity is mainly located between 5 and 10 km depth, in a really thin layer of about 200 m around Psathopyrgos fault (section P12, Fig.5e), but in a 2 km thick layer around Rion–Patras fault (section P16, Fig.5f), both gently dipping to the N–NW. The westernmost seismicity is a bit less well resolved at depth due to the network coverage, possibly explaining the larger thickness.

Moreover, we observe a seismic gap between the western part of the Marathias fault and the Mornos fault (Fig.3), at the transition between the ZW and the ZM. Either the seismicity occurs north of the Marathias fault, under the northern coast and follows the coastline, or it occurs along a E–W line in the continuity of the mapped-extension of the Trizonia–Mornos faults, likely following the Psathopyrgos fault plane at depth (sections P12 and 19, Fig.5).

3.2 Fault zone structures

3.2.1 The North Eratini zone (NEZ), a deep structure

The NEZ presents a specific seismic behaviour linked to the blind offshore fault which ruptured in 1995 (Bernard et al.1997; Lam-botte et al.2014). The deep seismicity located at more than 10 km depth under the northern coast defines the downdip extension of this low-angle rupture plane dipping north∼30◦(Bernard et al.1997; section P4, Fig.5a) . Moreover, according to Lambotte et al. (2014) the multiplets at the root of the fault show a similar dip around 30◦. Either the microseismicity occurs directly on the deeper part of the blind fault plane, which then could be longer than previously proposed, or the microseismicity nucleates at the edges of the fault plane and delineates its depth termination.

3.2.2 The transition zone (TZ), a structural boundary

The ZC and the ZW are clearly separated by a thin structure ori-ented N–S shallowly dipping north and containing numerous clus-ters (Fig.4). This discontinuity is not only visible in the seismicity

with the transition between a narrow part of the rift with shallow bathymetry further west (the Channel and the Delphic Plateau), and an older, wider, deeper part of the rift further east (the Mornos Canyon leading to the Corinth abyssal plain; Heezen et al.1966).

The integration of the 2008–2015 data to the 2000–2007 data set used in Lambotte et al. (2014) better highlights the discontinuity. The TZ is more clearly outlined by a north–south line of persis-tent clusters, which form a swelling of∼2 km high (Supplementary Information Fig. SC.1.b and c). This seismic alignment also delin-eates the western boundary of the 1995 earthquake rupture plane (Bernard et al.1997), where no or very little microseismicity has been recorded since. The origin and the nature of this structure is not yet well understood, but it has the same orientation as the Hellenic nappes (e.g Jolivet et al.2010) inherited from the past convergence during the Miocene.

3.2.3 The Trizonia Island zone (IZ), a second parallel structural boundary

As for the TZ, the Trizonia Island zone (IZ) shows a north–south alignment of several clusters, with sparse seismicity on each side (Fig.4). At depth, we observe that this deeper seismicity is contained in a thin layer, around 200 m thick, and is gently dipping north 15◦ (section P8, Fig.5d). On the contrary, the seismicity under the gulf is concentrated on a 1–2 km thick layer, almost flat, connecting to the root of the Fassouleika fault on the southern coast. The change from a thick flat layer under the gulf to the northern thinner layer of the IZ seems to coincide with the projection of Marathias fault plane at depth, dipping 55◦. The IZ mimics the TZ in terms of seismicity alignment, but does not seem to be associated to a fault system discontinuity.

3.2.4 The Aigion and Fassouleika fault zone (AFZ), a shallower activity

In the AFZ, the microseismicity clearly outlines the Aigion fault and the Fassouleika fault up to 3–4 km depth (section P7, Fig.5c). The dip of the structures is consistent with the one observed at the surface (50◦–60◦), also constrained by the AIG10 borehole (Cornet

et al.2004) which is crossing the Aigion fault at 760 m in depth, and is consistent with the geometry and the fault plane solutions of multiplets (Godano et al.2014; Duverger et al.2015). The faults are rooting in the dense seismic layer, 1–2 km thick, almost horizontal. The fault system is better outlined by the microseismicity over the whole 2000–2015 period than it has been shown in previous studies over a restricted observational span (e.g. Lambotte et al.

2014). Some kilometre-long structures are observed in the seismic layer but with an antithetic direction (zoom in section P7, Fig.5c), subparallel to the Trizonia fault plane extension. The Trizonia fault seems to mark the northern limit of the thick seismic layer, which further north becomes a really thinner layer aligned on a 15◦ dip plane as for the TZ and the IZ (section P7, Fig.5c).

3.2.5 The Rion–Patras and Psathopyrgos fault zone (RPZ), a connection between two gulfs

We have seen that the microseismicity in the ZM of the Corinth Rift could appear to be less dense, but this is actually the effect of the late widening of the network. The two main normal faults on the southern coast of the rift, the Rion–Patras and the Psathopyrgos faults, meet at the surface with a high angle change in strike. The observed microseismicity is mainly following these structures. The relocated clusters are aligned on an E–W straight line parallel to the Psathopyrgos fault, in the middle of the rift and then turn parallel to the Rion–Patras fault, following the coast line at N45◦E. We also observe some seismicity under the Rion–Patras fault, in its footwall, which is possibly connected to the inland A. Kastrisi fault or related to the Hellenic nappe stack (sections P15 and P16, Supplementary Information Fig. SC.1).

The sections P18 and P19 of the Fig.5, which are parallel to the Psathopyrgos and the Rion–Patras, respectively, present nice pic-tures of the seismicity occurring along the two fault planes. The known surface extension of the Psathopyrgos and the Rion–Patras faults are reported on the cross-sections. The seismicity is mainly located around 7 km depth in a layer varying from hundred-metre-thick to kilometre-hundred-metre-thick. The location of the southwesternmost seis-micity is not well defined due to the poor seismic network cover-age there, explaining the large dispersion in depth in section P18, (Fig.5g). We note a change in the seismicity at 17 km (horizontal axis) on the same cross-section which possibly images the junction between the two faults. The projection of this change at the surface does not fall on the Rion–Patras fault trace. However, in section P19, (Fig.5h), we clearly identify the seismicity extending along the Psathopyrgos fault which precisely corresponds with its mapped surface extension. In this section, the seismicity is concentrated in a very thin flat layer, 100 m thick, at 7 km depth and includes clusters. Further west and further east, the seismicity gets more scattered.

3.3 Spatio-temporal evolutions

3.3.1 Migration velocity during swarm activity

In the western Corinth Rift, during swarms, migrations of micro-seismicity have already been observed and some of them have been analysed in details. For instance, in 2001, the spatio-temporal evo-lution of the Agios-Ioannis earthquake swarm was modelled by Pacchiani & Lyon-Caen (2010) assuming a fluid-driven seismicity migrating upwards at a velocity of 0.02 km d−1. The 2003–2004 seismic crisis occurring along the Aigion and Fassouleika faults was associated to a hydroshear process caused by pore pressure mi-gration (Duverger et al.2015) at about 0.05 km d−1. In 2006–2007, another swarm with westward migration (Lambotte et al.2014) oc-curred in the middle of the gulf, close to Selianitika and Lambiri faults.

Here, we present some results on two large swarms which oc-curred in 2014, one in the ZM along the Psathopyrgos and Rion– Patras faults, and another in the middle of the gulf along the Kamarai fault system. During these seismic crises, a broad variety of migra-tion velocities is observed. All migramigra-tion velocities menmigra-tioned in the following text are projected along the E–W axis, which is the main migration direction at first order. True migration velocities are then equal or larger than the ones discussed here.

First, we describe the westernmost 2014–swarm. In July, we ob-serve a first slow westward migration of the microseismicity at about 0.08 km d−1, which started at the western end of the Psathopyrgos

fault (Fig.6). A second shallower eastward migration at a velocity of 0.3 km d−1occurred from mid-August along the Psathopyrgos fault. From mid-September, a new, faster, bilateral migration started close to the junction between the Rion–Patras fault and the Psathopyrgos fault, with several magnitude MW3 events. The eastward velocity

reaches 3 km d−1along the Psathopyrgos fault. Both westward and eastward velocities of this bilateral migration are of the order of a few km d−1. Contrary to previous swarm activities in this region, the two main faults seem to be connected and are activated as a continuous fault system during this crisis.

Second, we are interested in the 2014–swarm which developed from June to December along the Kamarai fault system (Fig.7). It started relatively deep, at 7–8 km depth, just after the 2014 June 8 MW4.1. This event is either located at the western end of the

Fassouleika fault, or on the Trizonia fault, the focal mechanism being compatible with both. Looking at the vertical cross-section (Fig.7b), the event is relocated in the continuity of the Trizonia fault extension at depth, leading us to favour a fault plane dipping south. Hypocentres of early aftershocks rapidly migrate to the north–east, over 3 km away from the main shock in 10 min. Knowing that a magnitude 4 event has a typical rupture dimension of about 1 km, this denotes a strong post-seismic effect. Then, the seismicity front migrated eastwards at 1 km d−1 for 2 more kilometres (Fig. 7e), while another slower migration at about 0.05 km d−1 continued during almost 3 months (Fig.7d). Two different classes of migration velocities, slow and fast, are observed there and this variability will be more discussed in Section 4.1.2. During the slow migration, the seismic activity jumped eastwards, with events a bit shallower (6 km depth) to finally come back deeper and fill in the spared area. From mid-August, another westward migration at 0.05 km d−1occurred along the Fassouleika fault extension and lasted more than 2 months. The seismic rate slowed down at the end of October, just before the occurrence of the 2014 November 7 MW4.7 event

located at 5.5 km depth, which triggered an intense and shallower swarm on the Aigion fault (Figs7a, c and d). Regular aftershocks are detected within hours on the shallow part (around 4 km depth) of the Aigion fault, with a lateral extension of about 8 km. An eastward migration of the aftershocks is observed at a velocity of 0.6 km d−1(Fig.7f). When the activity on this fault slows down, the Fassouleika fault remains active with the same westward migration at a velocity of 0.05 km d−1. The two faults seem to play alternatively and independently.

The seismicity from June to November starts below 7.5 km depth and migrates progressively upwards until 6.5 km depth, before the activation of the more superficial crisis in November, ranging from 3.5 to 7.0 km depth.

3.3.2 Relationship between shallow and deep seismicity

The last 2014–swarm showed that the shallowest activity along the Aigion–Fassouleika fault system occurred following a relatively shallow MW4.7 earthquake. Globally, over the 2000–2015 period

in the Corinth Rift, events shallower than 6 km depth mainly oc-cur in conjunction with the main swarm activities (Fig.8). We do not observe specific shift in time or pattern of the high activity of shallow events compared to deeper events or major events. We also often note that main shocks triggered shallower aftershocks either instantaneously or with a migration pattern as seen previ-ously. Moreover, we observe that the largest events (MW > 3) are

localized deeper than the smaller magnitude (Supplementary In-formation Fig. SC.2). At first glance, this characteristic might be

Figure 6. Seismic migration velocities during the westernmost (ZM) swarm in 2014. (a) Map of the seismicity migration along the Psathopyrgos, Nafpaktos and Rion–Patras faults. The longitude 21.6◦is 0 km distance. The black lines are the main normal faults. The colourbar represents the time in Julian days of 2014. (b) Temporal evolution of the microseismicity along the longitude. The circle size is proportional to the event magnitude. (c) and (d) are zooms in migrations represented by light grey frames in (b). Black lines represent the main migrations seen with the corresponding velocities in km d−1. The colourbar represents the event depth.

due to the relocation process, in that MW ∼ 3 events have a

dif-ferent frequency content than smaller events, thus are possibly less well correlated to them. However, relative location errors on the largest events are very similar to those of their neighbour events in the swarm, so we are confident in their relocated position. This observation suggests that faults are larger in the deeper part of the thick layer than in its more active central part. We suggest different interpretations: (1) this could reflect a geological heterogeneity or a lithological variation, with for instance a brittle or finely strat-ified limestone in upper part against a massive limestone in the deeper part; (2) it could be a sign of a mechanical evolution of the layer, with an intense fracturing of the medium linked to the root-ing normal faults. The shear stress layer would be more fractured with smaller structures in the upper part than in the deeper part, where the fault planes are better preserved with a higher magnitude potential.

3.3.3 Long-term and short-term multiplets and repeaters

We map all multiplets composed by at least 10 relocated events, which represents more than 1000 families (Fig.9). The first obser-vation is that multiplets are localized in dense seismic area (Sup-plementary Information Fig. SC.1). The lower number of multiplets found in the ZM is likely related to the recent installation of seismic stations there.

Some of the multiplets are clustered in time, others last several years (Fig. 9b). Godano et al. (2015) noted that multiplets with a regular or persistent activity are located under or close to the northern coast of the Corinth Gulf, at the root of the 1995–fault. We therefore analyse the temporal activity of the major multiplets in order to check these observations over the whole time period (16 yr). We classified the multiplets into two groups: long-term and short-term. The long-term multiplets are regularly activated throughout the observational period or for at least several years.

Figure 7. Seismic migration velocities during the western (ZW) swarm in 2014. (a) Map of the seismicity migration along the Selianitika–Fassouleika–Aigion fault system. The longitude 21.6◦is 0 km distance. Black lines are the main normal faults. The colourbar represents the time in Julian days of 2014. Focal mechanisms are from NOA (National Observatory of Athens,1997). Red lines delineate the two cross-sections displayed in (b) and (c) with their lateral extent. (b) N–S cross-section of the events occurring at the beginning of the crisis. (c) N–S cross-section parallel to the first one, further east, of the events occurring later during the crisis. (d) Temporal evolution of the microseismicity along the longitude. Circle size is proportional to the event magnitude. (e) and (f) are zooms in specific migrations, represented by light grey frames in (d). Black lines represent the main migrations seen with the corresponding velocities in km d−1. The colourbar represents the event depth.

Almost all the multiplets in the ZC and along the TZ belong to the long-term ones, that supports Godano et al. (2015)’ observations. Multiplets in the IZ, as well as the ones located at the border of the silent patch between the Mornos fault and the Marathias fault, are also long-term multiplets. On the other hand, the short-term or burst-type multiplets appear only once and do not activate again later, even during other swarms in the same area. In the AFZ, multiplets mainly belong to this burst-type. They last less than 1 yr and often occur during seismic crises or during aftershock sequences with no further

reactivation during next swarms. Furthermore, on cross-sections of Fig.9, we can see that long-term multiplets are systematically deeper than the short-term ones.

To complete this temporal analysis, we computed the interevent times tr inside each repeater to characterize the recurrence of

repeater-events (Fig.10). We remind that repeaters have been ex-tracted using a similarity threshold Sc= 0.9 for a 5 s window

en-compassing the P and S waves together. 30 per cent of events from the database and 45 per cent of multiplet events are inside repeaters.

Figure 8. Histograms of seismic activity in time at different depth ranges. (a) Histogram of the number (in log scale) of superficial events (0≤ depth ≤ 6 km) through time. (b) Histogram of the number (in log scale) of deeper events (6≤ depth ≤ 12 km). Yellow histograms correspond to events of MW≥ 3.0. Vertical red lines indicate the occurrence time of the largest events (MW≥ 3.8). The line is plain if the event occurred at the corresponding depth range, it is dashed otherwise. (c) Histogram of the total number of CRL events in the first 15 km depth. Main swarms are indicated in grey. Time bins are computed every month. The CRL network was enlarged in 2010, which explains the increase of the number of events from this year.

Figure 9. Maps of the multiplets. Each circle represents a multiplet having more than 10 events, which is located at the mean hypocentral position of its relocated events. Colours represent the duration (in seconds) of multiplets in log scale. White vertical lines in the colourbar indicate several temporal points of reference: 1 hr, 1 d, 1 month, 1 yr and 10 yr. The same map with colours corresponding to the number of events per multiplet is in Supplementary Information Fig. SC.3.

Figure 10. Analysis of the interevent times of repeaters with at least five events. (a) Number of events as a function of the mean interevent time T0(in seconds) of repeaters. Colourbar corresponds to the duration (in seconds) of repeaters in log scale. Coloured dashed lines are constant duration. Repeaters with small interevent times (tr≤ 2 hr) never last long and never have a large number of events. (b) Histogram of repeaters’ mean interevent time (in seconds). (c) Histogram of repeaters’ duration (in seconds). (d) Stacked probability densities of normalized interevent times (tr/T0) for repeaters. Green dashed line is the fitted tr/T0 gamma-law decay estimated thanks to the smallest normalized interevent times (green dots). Grey dashed line represents the number of samples equal to 1. The further we are from this line, the better the results are since more observations are included in the estimation. We plotted the differences between data and the regressed law in log scale and we estimated the standard deviation (dotted green lines) of the data to the law. Red dots highlight a small bump just above 1σ for 10−1< tr/T0< 100, corresponding to some quasi-periodic events. (e) Histogram depicting the coefficient of variation (COV) distribution of repeaters.

For the presentation of the results, we only kept repeaters with at least five relocated events (Fig.11). It represents more than 1000 families. We used the periodicity definition of Lenglin´e & Marsan (2009) and computed the mean interevent time (T0) as the mean of

the time intervals between successive events of a repeater. Long-term repeaters have higher mean interevent times than short-term repeaters (Fig.10a). For a same number of events per family, this relation is trivial but in the general case, it is not. The important feature is the dispersion of the interevent times for a given number of events per repeater. We observe a lack of large repeaters (N≥ 10 events) with low mean interevent time (T0≤ 2 hr).

The bimodal distribution of the repeaters’ mean interevent time and duration (Figs10b-c) is likely associated to their spatial loca-tion. The hole in the distribution at around respectively 2 months for interevent time and 8 months for duration, is probably a signature of the return period of a swarm in the same area. A swarm tends to last some months, but is never active more than 8 months. The time span of short-term repeaters is limited to the occurrence of specific seismic crisis, whereas long-term repeaters span over sev-eral swarms and periods of lower background seismicity, separated by several months, even years, in a same region.

For each repeater, we defined the normalized interevent times as tr/T0, which are then sorted by time bins over all families. To

obtain a probability density, the number of recurrence times falling

in each time bins is divided by the length of the time bin and by the total number of normalized interevent times. For this, we consider a gamma distribution: ρ(tr)= A  tr T0 _{γ −1} exp  −tr T0  (1) where A is a constant (Hainzl et al.2006). Ifγ = 1, then the seis-mic activity follows a Poisson law, that is, a random occurrence of events with no interactions (background seismicity). These events occur at a known mean frequency independently of the elapsed time since the previous event. On the other hand, ifγ = 0, the seismic activity follows an Omori law (Omori1894) expressing the delayed interactions between a mechanical disturbance gen-erated by an earthquake and the seismic rupture of the nearby faults. In other words, it characterizes the mainshock–aftershock sequences. Finally, if 0 < γ < 1, then the seismic activity is clustered with a high number of events in a given time without particular interactions, and ifγ > 1, it follows a regular seismic cycle.

We represent the probability density function of the normalized interevent times and fitted the data by a gamma-law decay (Fig.10d). We findγ ≈ 0.35, typical of cluster activity, and observed a decrease similar to an Omori-law decay. We also note a really small bump

Figure 11. Periodicity of repeaters. (a) Map and north–south, east–west cross-sections of repeaters having more than five events. Each circle represents a repeater, which is located at the mean hypocentre position of its relocated events. The circle size is proportional to the number of events in the repeater. The colours indicate the coefficient of variation (COV) of the interevent times for each repeater. A COV equal to 0 means that the repeater is fully periodic. A COV equal to 1 means that the repeater interevent times are random. A larger COV corresponds to repeaters having some much longer and some much shorter interevent times than its mean interevent time. (b) Temporal occurrences of events of some quasi-periodic repeaters. The colour and the size of circles are proportional to the event magnitude. These particular repeaters are identified by stars and numbers on the map and cross-sections.

between 10−1 and 100 _{s with a factor of∼1.25 (0.1 in log scale}

difference) compared to the linear log-log decay. The dynamics of earthquakes seems then dominated by interactions between asper-ities rather than steady stress accumulation on a single asperity. We can distinguish three different regimes: (1) about 65 per cent of the events are in clusters (temporal interactions between events of

Omori-type), (2) less than 35 per cent follows a Poisson law (ran-dom background seismicity) and (3) the remaining events occur quasi-periodically (red frame in Fig.10d).

We further investigate the data in order to potentially identify the periodic or quasi-periodic earthquake recurrences, typical of repeater behaviour. To do so, we compute coefficients of variation

(COV) as the standard deviation of the interevent times divided by the mean interevent time by repeater. A COV of 0 implies a perfectly periodic recurrence, whereas a COV of 1 is characteristic of a Poisson sequence. The majority of COV values are between 0.5 and 1 (Fig. 10e). This is consistent with the fitted gamma-law decay seen in the probability density function. However, few repeaters show a very low COV value, which is expected in case of periodicity. We then plot some sequences of these repeating events in Fig.11. A concern about distributions of interevent times is that chances are that some otherwise similar events may actually be missing (not detected/located) or not included in a repeater due to the strict criteria for waveform similarity. Neverthless, this repeating behaviour can be interpreted as the result of a constant stressing rate acting on the asperities (Lenglin´e & Marsan2009) and could suggest aseismic slip. The families last generally several years, but we also observed one example of a really short-term repeater, lasting less than half a day with a COV equal to 0.2. Nevertheless, low COV– value repeaters correspond mainly to the long-term repeaters. They are concentrated in the same zones as seen before, that is to say the ZC, along the TZ and the IZ, and around the patch with no seismicity between the Mornos and the Marathias faults. On the other hand, the AFZ concentrates many high COV–value repeaters, as well as the 2001–cluster and the 2013–cluster further south.

The events of repeaters should have similar moment magnitudes, and there must exist a relation between the mean recurrence time and the mean moment of a repeating sequence. To test this assumption, we computed the mean moment magnitude M0of each repeater and

compared it with its mean recurrence time T0. For the whole repeater

population (Fig.12a), no trend was highlighted and long interevent times spread over the same magnitude range as the short ones. However, by selecting the most periodic repeaters (COV< 0.6), we show that they primarily have large T0and that these repeaters with

large T0seem to follow a linear relationship as Nadeau & Johnson

(1998)’s one (Fig.12b). However, the magnitude range explored in our case is very narrow (only one order of magnitude) and too small to push further the interpretation in terms of mechanism.

We also tested two classical models of seismic hazard: (1) the time-predictable model, for which the rupture threshold is known but not the stress drop and (2) the slip predictable model, for which the stress drop is known, not the rupture threshold. For consecutive events of a repeater, we compared then either the magnitude of the first event to the interevent time or the magnitude of the second event to the same interevent time (Supplementary Information Fig. SC.4). However, no valuable information has been emphasized from these tests. We thus constrain this analysis for only the quasi-periodic repeaters (COV≤ 0.4). This sub-group of repeaters reflects the magnitude distribution of the whole repeater population but not the interevent time distribution. Recurrence times of the quasi-periodic repeaters are large, 70 per cent of them are larger than 107_{s, that is,}

∼ 4 months. Moreover, for this population of large interevent times, we observe a trend consisting in an increase of the delays with an increase of the magnitudes of either the first or second event. This population is probably composed of real repeaters in the sense that a same asperity breaks, more or less regularly.

4 D I S C U S S I O N

The relocation of this large data set of more than 200 000 earth-quakes was challenging because of the intense computational task required for the analysis of waveforms and parametric data. To ac-complish the processing, we had to make choices such as dividing

the region in several areas, deciding the distance threshold for cross-correlations, determining damping values for the double-difference relocations and selecting a dissimilarity cut-off for multiplet classi-fication. Although our results could probably be improved by study-ing limited areas individually Pacchiani & Lyon-Caen (2010); Du-verger et al. (2015); Kapetanidis et al. (2015) and by optimizing configuration parameters for each processing step, the results of this study provide a uniform view across the whole western Corinth Rift. Thus, this new, precisely located catalogue of events can be used to analyse spatial and temporal seismicity patterns across the whole highly active ZW.

4.1 Seismic activity along the southern fault systems

4.1.1 Active faults

Our relocation shows that the southernmost faults (West and East Helike, Kerinitis, Pirgaki) produce some microseismicity mostly in their downdip continuation, despite the reduction of their slip rate related to the northward migration of fault activities (Goldswor-thy & Jackson 2001; Palyvos et al. 2005; Ford et al. 2013), which favours the activity of the younger Kamarai fault system (Aigion–Fassouleika–Selianitika–Lambiri), as inferred from geo-logical (Flott´e2003) and GPS observations (Avallone et al.2004). In particular, the 2001–swarm occurred on the Kerinitis faults (Pac-chiani & Lyon-Caen2010; Lambotte et al.2014), and the 2013– swarm is likely to be associated to the Pirgaki fault (Kapetanidis

et al.2015), although both faults are supposed to be geologically inactive. This suggests that these deep swarms are driven by pore pressure transients, and not directly by the tectonic shear loading rate. The large extension rate of the rift may, however, play a role in keeping a high permeability in the faulted volume at the intersection between the fault roots and the brittle geological layer at depth.

Further west, some microseismicity is located in the footwall of the Rion–Patras fault, which could be either associated to the inland Kastrisi fault, or to a weak geological layer, possibly the Phyllades nappe beneath the Tripolitsa nappe (Fig.1) outcropping north to the Rion sill on the northern coast, and 20 km to the south of Patras (Beckers et al.2015).

4.1.2 Driving mechanisms in the middle of the rift

The seismicity is concentrated in 1–2 km thick active layer in the middle of the rift, containing numerous multiplets with a wide range of strikes and dips (Lambotte et al.2014; Godano et al. 2015). This layer has a high variability of internal structures along the rift over short distances. The best-fit planar structure to this thick microseismic layer is a plane roughly oriented N260◦E, dipping 5–10◦N.

Along the Kamarai fault system, characterized by a high concen-tration of burst–like multiplets, we systematically observe seismic migration velocities of the order of tens or hundreds metres per day during seismic swarms. This is for instance the case during the 2014–swarm which has shown two migrations at 0.05 km d−1 in between the aftershock sequences. This matches the observations and interpretations by Duverger et al. (2015) during the 2003–2004 seismic crisis, and the computed migration velocities compare with that of a pore pressure front diffusion (e.g. Chen et al.2012).

On the Aigion fault, which roots into this seismically active layer at about 6 km depth, we detect a notable seismicity up to a depth (∼3 km) shallower than previously observed (Lambotte et al.2014).

Figure 12. Variation of the mean recurrence time of a repeater as a function of the mean seismic moment of its events. (a) Graphs for all repeaters. Each circle corresponds to a repeater. Circle size is proportional to the number of events contained in the repeater. Colourbar refers to the repeater COV. Black dashed line is the log(T0)= 4.85 + 0.17 log(M0) relation reported by Nadeau & Johnson (1998), where T0is the mean recurrence time in seconds and M0is the mean seismic moment of the repeater in dyn cm (1 N m=107dyn cm). Histograms count the number of repeaters through mean moment magnitude bins (up), and mean interevent time bins (right). (b) is the same as (a) for a sub-population of more periodic repeaters (COV≤ 0.6). Red dashed line illustrates a positive trend among repeaters with large interevent times and corresponds to the equation log(T0)= 4.4 + 0.17 log(M0). Only the intercept differs from Nadeau & Johnson (1998)’s relation.

It may indicate that the deeper part of the fault is continuously creep-ing over years while the upper part is partly locked, but can also par-tially creep, either continuously or episodically. This interpretation can be supported by InSAR observations (Elias2013), which show creep on the Aigion fault with an uplift rate of about 2 mm yr−1.

4.1.3 Mechanics of short-lived repeaters

In the middle of the gulf, along the Kamarai fault system, the as-perities of repeaters are excited for a short time, usually during a single swarm, and then disappear, never break again. This phenom-ena could be interpreted as (1) a gradual abrasion of the roughness of the asperity leading to a loss of the seismic behaviour of the as-perity, (2) the mechanical locking of the asas-perity, (3) the complete release of the available elastic energy and/or (4) the progressive alteration of the large scale rock mass resulting into a decorrelation of the waveforms.

The first hypothesis assumes that each repetitive rupture smooths the fault surface and reduces its friction, until it becomes aseismic and can slip continuously. This would imply a very speculative difference in the abrasion properties of these asperities with respect to those producing long-term repeaters.

The second hypothesis proposes a locking of the fault which may arise from a dislocation step cutting the fault surface due to slip on an intersecting fault. This seems plausible, as many small faults are expected to slip, seismically or aseismically, in the over-pressurized and sheared rock mass activated during a swarm episode.

The third hypothesis is based on the energy accumulation and release. An easy way to stop a multiplet activity is to reduce the local loading stress by the generation of a rather large slip, possibly aseismic, on a neighbouring fault, occurring during the swarm. But it is unlikely that such persistent stress shadow would systematically lock all short-term multiplets appearing in all swarms. Alternatively, the stored energy may be simply exhausted by the multiplet activity itself. In this case, the constant shearing rate due to the long-term tectonics, about 10−6yr−1, is too small to allow fast reloading in the

duration of the swarm. Only slow slip around the asperities could amplify the local shear strain enough for allowing several seismic ruptures until reaching a total shear stress release. One might also consider an increase of the pore pressure leading to a decrease of the effective normal stress, which could trigger the rupture. In this case, five events with stress drop of about 1 MPa would require an overpressure of 8 MPa to be generated, assuming a friction coefficient of 0.6. Such high value is common in fluid injection experiments (e.g. Cornet et al. 1997), but probably not realistic for natural fluid pressure perturbations. The slow slip mechanism loading is thus the most likely dominant model under the hypothesis of exhaustion of the stored elastic energy.

The last hypothesis for the repeater disappearance is the possible decorrelation of the waveforms due to seismic velocity changes around the fault zone. The stress perturbations caused by the swarm activity have to be strong enough to sufficiently change the local velocity structure and significantly modify the waveforms generated by a single asperity. Moreover, the persistence of other repeaters from deeper zones imply that these modifications, if any, have to be limited to a few kilometres in the vicinity of the clusters, implying very large structural perturbations of the geological layer, which we believe are unlikely.

To conclude, the locking model and the exhaustion model as-sisted by slow slip are our preferred explanations for the short life of swarm-related repeaters. Note that in the case of non-repeater multiplets, that is, involving neighbouring rupture areas instead of a single asperity, the most likely explanation is the simple exhaustion model, as the locking model would require to separately operate on each of the neighbouring asperities.

4.2 The 1995–fault rupture

The observation of deep events at almost 15 km depth in the downdip continuation of the 1995 blind rupture plane could be interpreted as a longer fault extension at depth than previously established. The relatively sparse distribution of events deeper than 10 km may be